On October 12, NVIDIA announced that it had contributed key rack and compute tray designs for its next-generation AI platform, Vera Rubin, to the Open Compute Project (OCP). Making GPUs more powerful does not by itself increase computing capacity. The enclosures and liquid-cooling equipment that house them must also be built in volume, connected to power and brought into service. The MGX 1.2 specification standardizes a rack design that supports up to 250kW, along with plug-in liquid-cooling connections. It shows how much of the design companies that build AI servers and companies that deploy them can share, and where each side must still validate its own setup.
From Blackwell to Vera Rubin: the design contributions continue
NVIDIA's announcement explains that, through its modular architecture "MGX," it has contributed the key rack and compute tray designs for Vera Rubin to the OCP. A compute tray is the enclosure that holds CPUs and GPUs and slides into the rack. Contributing these designs to the OCP lets the manufacturers who build such equipment work from a common design.
Looking back at how the designs were released makes the novelty of this announcement easier to grasp. NVIDIA had already contributed rack-related designs to the OCP for the Blackwell generation, and last year it previewed improvements for Vera Rubin.
| NVIDIA announcement date | Design covered | Stage indicated in the announcement |
|---|---|---|
| October 15, 2024 | GB200 NVL72 rack structure, tray mechanics, liquid-cooling and thermal requirements | Key electrical and mechanical designs contributed to the OCP |
| October 13, 2025 | Improved rack and compute tray for Vera Rubin | Contribution of the designs to the OCP previewed |
| October 12, 2026 | Key Vera Rubin rack and compute tray designs | Announced as contributed to the OCP through MGX |
What has changed is not the debut of MGX but the OCP contribution of the Vera Rubin rack and tray designs previewed in 2025. The table compares the generation covered and the stage of contribution in each year's announcement. It does not show when OCP review was completed or when all design files were made public.
MGX 1.2 standardizes the power and cooling connection specifications and adds configurable I/O bays and support for 19-inch equipment. If equipment configurations can vary within a common specification, manufacturers can more easily avoid rebuilding enclosures and connectors for each customer.
However, what was disclosed this time is the key rack and tray designs. It is not an announcement that opens up the internal circuitry of the Rubin GPU or semiconductor manufacturing technology. Aiming for interoperability under a common specification is also not the same as saying that parts from different manufacturers can be swapped unconditionally.
Hoses and fans can go because liquid cooling is rebuilt around the connection method
According to the DGX Vera Rubin NVL72 user guide, the rack holds 18 compute trays, each carrying two Vera CPUs and four Rubin GPUs. Nine NVLink switch trays link the 72 GPUs in total. At this scale, the procedure for assembling and replacing a single tray also affects manufacturing and maintenance of the entire rack.
NVIDIA's technical explanation says the Vera Rubin compute tray was redesigned to eliminate cables, hoses and fans by using a midplane board that connects the internal components. Liquid cooling has not been abandoned. Coolant flows from the rack's piping to the cold plates attached to the CPUs and GPUs.
What MGX 1.2 adds is a liquid-cooling method that connects when the tray is slid into position. NVIDIA also cites MQD and MQDB as related liquid-cooling technologies. Including the cooling-component connections in the common design is intended to simplify assembly and on-site replacement.
Instead of routing cables and hoses one by one, components are combined through a board and defined connection points. That is where the process changes. Manual work can be reduced, but validation remains: the precision of the insertion position, the quality of the connection and whether the required volume of coolant flows. Ease of assembly comes from matching the connectors to the rack-side design.
Be careful about what the "no cables, no fans" description covers. The copper cabling that links trays over NVLink remains in a cartridge at the back of the rack, and the user guide's power shelf section also lists air-cooled power supply units. What is eliminated is the wiring and fans inside the compute tray, not all wiring and airflow for the entire rack.
The power delivery section is liquid-cooled as well. MGX 1.2 includes a liquid-cooled busbar that supports 5,000A of current. A busbar is a strip-shaped conductor that carries large currents. Delivering large amounts of power to GPUs requires dealing not only with the heat from the chips but also with the heat generated in the conductors that carry that power.
Reading the 250kW rack and 800V distribution separately
The "up to 250kW" that NVIDIA cites for MGX 1.2 is the range of rack deployment power that the architecture supports. It is not a measurement showing that every Vera Rubin NVL72 product constantly draws 250kW. Alongside 5,000A and 800V, it can look like a single power specification, but each figure describes a different subject.
| Figure | What it represents | Location or configuration it applies to |
|---|---|---|
| Up to 250kW | Range of rack deployment power supported by MGX 1.2 | The rack's design-level capability |
| 5,000A | Current supported by the liquid-cooled busbar | Conductors inside the rack |
| 54V DC | DC voltage converted and supplied by the power shelves | In-rack power delivery described in the DGX Vera Rubin NVL72 user guide |
| 800V DC | Distribution voltage for high-density AI facilities | Configurations using facility power paths or dedicated power racks |
The rack capability and conductor current are based on this MGX announcement, the 54V figure on the DGX power shelf description, and the 800V figure on NVIDIA's explanation of its power distribution design. Multiplying these figures together does not yield the power consumption or guaranteed rating of any individual product.
The aim of 800V DC is to change the path that carries power from the facility to the computers. Distributing direct current at a higher voltage reduces the burden on conversion equipment and wiring. NVIDIA also describes a migration approach that adds dedicated 800V DC power racks to existing AC facilities, so it does not assume only configurations that convert an entire building to DC at once.
The "K-1" shown by Visionbay this time is also a plan for an AI facility using 800V DC. It takes the approach of concentrating AC-to-DC conversion in cluster-level power equipment and introducing it step by step, starting where the benefit is greatest. The common rack design and facility-side power distribution upgrades are related, but this is not an announcement that every current DGX configuration has switched to 800V input.
Even with support for 19-inch equipment and common connection points, the existing power and cooling capacity at a site does not automatically increase. Whether a deployment site can deliver the required power and carry away the heat generated has to be confirmed facility by facility.
The value of a common design extends from manufacturing to operation
Racks filled only with Vera CPUs also share the design principles of third-generation MGX. While Vera Rubin NVL72 handles large-scale training and inference, the CPU-side equipment supports tool calls by AI agents, coordination of processing and execution in isolated environments. The idea is to handle machines with different roles in the same AI data center using a common power and cooling design.
Lenovo's product description for OCP 2026 shows concretely where the manufacturer's work remains. For its "AI Cloud Gigafactory" using Vera Rubin, the company says it provides equipment integration and validation in addition to its own design and manufacturing. It also combines its liquid-cooling experience with Neptune and post-deployment operational support.
As common designs spread, manufacturers can reduce the burden of deciding basic connection methods from scratch. What then distinguishes products and services is how equipment is fitted into a customer's facility, how the group of machines is validated and how recovery is handled after a failure. Lenovo's description presents a concept of taking on this role; it is not the result of an independent test confirming reductions in deployment time or cost.
NVIDIA is also extending its scope to the process of running the finished racks. The open-source operations software added to DSX OS this time includes Config Manager, which aligns switch settings, and Cluster Readiness Engine, which checks GPU groups with workloads close to real ones. Rack Management Service handles equipment and firmware management at the rack level, and NVIDIA also released a foundation that supports deploying physical servers and adding users.
Handing out enclosure blueprints alone will not run a training job. Networks must be configured, equipment must be checked for its ability to withstand load and its state must be managed during operation. The move to pair hardware standardization with operations software suggests an aim to standardize, too, the process of turning delivered racks into usable computing facilities.
On the facility side, NVIDIA also announced Open Data Center Specification v0.7, covering power, cooling and layout, with AMD, Google and Meta. The OCP Grid Resilience Specification, announced with Google, Meta and Microsoft, addresses how systems respond when the power grid is disturbed. Designs that make the inside of the rack easier to build are advancing in parallel with designs that connect to facilities and the power grid.
This opening-up could also make NVIDIA's platform easier to adopt. If component makers build products targeting the same specification, it becomes easier to increase the number of suppliers deploying companies can choose from. At the same time, that supply chain would also support NVIDIA's rack designs. Sharing designs and encouraging broad use of its platform are compatible.
The measures for judging future adoption are whether equipment built on the common design passes validation in customers' facilities, connects with operations software and can be supported through replacement and recovery. If those conditions are met, it should become easier to standardize the steps that carry Vera Rubin's computing power from the manufacturer's factory to a data center where jobs actually run.